# ComfyUI workflow for the Qwen3-0.6B adapter

This folder contains a ComfyUI workflow that drives **FLUX.2-klein-4B** with the
**Qwen3-0.6B + adapter** text encoder instead of its native Qwen3-4B encoder.

- Workflow: [`flux2_klein_qwen3_06b_adapter.json`](flux2_klein_qwen3_06b_adapter.json)
- Custom node (the two `KleinAdapter` nodes):
  [github.com/recoilme/klein-qwen3-adapter-comfyui](https://github.com/recoilme/klein-qwen3-adapter-comfyui)

## Quickstart (from scratch)

From an empty machine to a first image — verified end to end on ComfyUI 0.36
(~8.4 GiB of downloads, ~1.4 GiB more on the first run):

```bash
# 1. ComfyUI itself — skip if you already have one
git clone https://github.com/comfyanonymous/ComfyUI
cd ComfyUI
python -m venv .venv && . .venv/bin/activate       # Python 3.10+
pip install -r requirements.txt

# 2. the custom node that provides the two KleinAdapter nodes
cd custom_nodes
git clone https://github.com/recoilme/klein-qwen3-adapter-comfyui
cd ..

# 3. models (paths are relative to ComfyUI/)
mkdir -p models/klein_adapter
curl -L -o models/diffusion_models/flux-2-klein-4b.safetensors \
  https://huggingface.co/Comfy-Org/flux2-klein/resolve/main/split_files/diffusion_models/flux-2-klein-4b.safetensors
curl -L -o models/vae/flux2-vae.safetensors \
  https://huggingface.co/Comfy-Org/flux2-dev/resolve/main/split_files/vae/flux2-vae.safetensors
curl -L -o models/klein_adapter/adapter_v14_bal.safetensors \
  https://huggingface.co/AiArtLab/qwen3-0.6b-4b-adapter/resolve/main/adapter_v14_bal.safetensors
# Qwen3-0.6B (1.4 GiB) needs no manual step: it is pulled from the Hub on the first
# run. Offline machine: fetch it beforehand with `hf download Qwen/Qwen3-0.6B`.

# 4. start ComfyUI
python main.py                                      # then open the URL it prints (:8188)

# 5. in the UI: Workflow -> Open ->
#    custom_nodes/klein-qwen3-adapter-comfyui/workflows/flux2_klein_qwen3_06b_adapter.json
#    (or drag this folder's JSON onto the canvas), type a prompt, press Run.
```

768×1280 with the distilled model takes ~2 s per image on an RTX 5090 and peaks
at ~12.6 GiB of VRAM (mostly the klein DiT itself).

## Models

```
ComfyUI/models/
├── diffusion_models/
│   └── flux-2-klein-4b.safetensors        # distilled klein DiT (stock), 7.2 GiB
├── vae/
│   └── flux2-vae.safetensors              # stock Flux2 VAE (untouched), 321 MiB
└── klein_adapter/                          # created automatically by the node
    └── adapter_v14_bal.safetensors        # this adapter, 840 MiB
```

- `flux-2-klein-4b.safetensors`: the klein DiT in ComfyUI format, from
  [Comfy-Org/flux2-klein](https://huggingface.co/Comfy-Org/flux2-klein)
  (`split_files/diffusion_models/`). The base model `flux-2-klein-base-4b.safetensors`
  works too (see the settings below).
- `flux2-vae.safetensors`: from [Comfy-Org/flux2-dev](https://huggingface.co/Comfy-Org/flux2-dev)
  (`split_files/vae/`).
- `adapter_v14_bal.safetensors`: from this repo (see above).
- Qwen3-0.6B is downloaded automatically from HF on first use.

## Run

1. Load `flux2_klein_qwen3_06b_adapter.json` in ComfyUI.
2. Set your prompt in the **Positive prompt** node.
3. Queue.

Settings in the workflow: distilled klein, **4 steps, guidance 1.0, euler**,
768×1280. For the base model use `flux-2-klein-base-4b.safetensors`, 50 steps,
guidance 4.0.

## Notes

The node computes the text conditioning identically to `example.py` in this
repo (same chat template, same layer taps `2,9,14,18,23,27`, same fp32 adapter,
same `drop_first=5`) — checked numerically, the two tensors are bit-equal
(`(1, 251, 7680)`, max abs diff 0.0). The image itself is then produced by
ComfyUI's own sampler (its noise, scheduler, text-position ids and 512-token
padding of the conditioning), so it is not pixel-identical to `diffusers` — the
same framework difference you'd see with the native klein encoder.